role
Lead designer responsible for defining the interaction model and behavior for the selling experience, from concept through high-fidelity design.
timeline
focus
system
tool
C
ontext
As Zillow's seller experience evolved, it became clear that the surface architecture itself was working against homeowners. The off-market home details page (HDP) and the owner dashboard overlapped in function, with no clear purpose defined for either, leaving homeowners with no reliable way to know which to use for what.
The seller Plan Tab was being introduced as a fourth surface, and the team wanted to add Voyager, Zillow's conversational AI, on top
of that. Before any of it could scale, the architecture itself needed a clear answer for what each surface was for and how they related to one another.
The Homeowner and Seller Surfaces POV emerged from this context, a project I led to resolve that structural problem. AI Seller was the first piece of that POV that moved into execution, specifically the Plan Tab portion, with the purpose of becoming Zillow's flagship agentic AI experience for sellers.
At the product level, AI Seller was a bet on whether an AI assistant could meaningfully carry someone through one of the most complex, high-stakes journeys in the product, selling a home, without overwhelming them or eroding their trust. Voyager itself existed only as a desktop chat experience for searching homes, with no established behavior model for how it should guide someone through a structured, non-linear process on mobile.
Because this was meant to be the flagship agentic AI product for both the seller vertical and Plan Tab broadly, the risk wasn't any single screen. It was the pattern, keeping chat from competing with the canvas it was meant to support, and the behavior, how much the AI should drive versus how much control the seller kept. Getting either wrong at this stage would compound across every feature built on top of it.
Problem
An audit showed the off-market home details page (HDP) and the owner dashboard overlapped with no clear purpose for either. My POV was to sunset the dashboard and consolidate into a single HDP.
That gave each surface a distinct job, the HDP for viewing home value, comps, and equity, plus other homeowner tools, with the ability to update the home details that affect that value; Plan Tab for planning and executing a move or renovation; and Voyager as the explainer and guidance layer, connecting the two based on user behavior.
In a design sprint I co-led with my product partner to validate that POV, I explored concepts ranging from lightly enhancing the existing tools to a fully conversational model, an idea that took shape after I read about a seller who'd sold his home almost entirely through an AI chat. That direction gained traction with leadership and confirmed Plan Tab as the execution surface for decision-making and the selling process.
Design
approach
I started mobile first because mobile was the harder problem given the constrained space. Early exploration stayed scoped to the existing Plan Tab framework, but stakeholders pushed back, they weren't satisfied with that framework and didn't want us beholden to its constraints.
Stakeholders pushed for more, a full vision rather than something scoped and safe. That led to a combined design sprint with the Plan Tab team, who were doing their own vision work in parallel. We generated a set of distinct concepts, reviewed them with stakeholders, and narrowed to three for user testing, centered on how Voyager should behave, since it had never existed on mobile before.
My prototype's information architecture came out as the clear winner from that research effort and became the interaction foundation for MVP. Due to a tight timeline, my teammate and I split the first milestone. I took net proceeds, the more ambiguous flow. She took home value, and I was also responsible for defining and fine tuning the Voyager AI patterns and behaviors that both flows inherited.
Scaling the vision
Part of the interaction framework was defining what guidance meant, because it drove the behavior. It was built on the principle that control is what builds trust, a finding that came through consistently in the research. That came down to three facets, ambient guidance, chat for depth, and next-best-step guidance.
Design
intent
AMBIENT CONTEXTUAL GUIDANCE
NEXT-BEST-STEP GUIDANCE
Contextual insight cards connect the data to the job to be done and the seller's personal situation. They're also an ingress into chat, for going deeper. This meant sellers who never opened chat could still get the same insights, so value wasn't held solely inside the conversation.
Solution
A short onboarding opens the experience, asking for the seller's timeline to sell, then their most pressing question about the process, before displaying relevant content to gather feedback. That sharpens what Voyager knows and helps determine a seller's first step in the journey.
Onboarding ends by landing a seller on the milestone page itself, with a dismissible ambient card surfacing the suggested next step. They can act on it directly, or dismiss it to see the full milestone journey end to end.My prototype's information architecture came out as the clear winner from that research effort and became the interaction foundation for MVP. Due to a tight timeline, my teammate and I split the first milestone. I took net proceeds, the more ambiguous flow. She took home value, and I was also responsible for defining and fine tuning the Voyager AI patterns and behaviors that both flows inherited.
This is where a seller lands on repeat visits. Because selling isn't linear, the current focus is surfaced first, along with progress and any new insight picked up since their last visit.
CONTEXT-FIRST ONBOARDING
A dismissible welcome pattern introduces how Voyager can assist on this flow, like walking through costs to sharpen the payout estimate. Voyager can guide each step and navigate to the next best step, but a seller can just as easily use the tab bar to move through it themselves. To move to a different milestone, a seller selects the back button, returning to the milestone page.
The net proceeds card shows what a seller would walk away with today and the drivers behind that number. Below it is an ambient insight card explaining those drivers, along with an ingress into Voyager that triggers contextual prompt chips specific to that insight, for a more intelligent experience.
Below that is the list price tool, then the cost breakdown in dollar and percentage terms, followed by an ambient insight card on the biggest cost levers. The next-best-step is always visible at the bottom, so sellers who don't engage with chat can still see what their next step should be.
NET PROCEEDS OVERVIEW
The list price tool is an interactive slider defaulting to the Zestimate, with the ability to adjust it ±10%. As a seller moves away from the suggested price, a status tag escalates from a lighter caution to a warning, giving them a read on where they stand.
Moving the slider updates the timestamp, and recalculates the payout and cost estimates accordingly. The tool also adjusts for Zestimate changes while the seller is offline, keeping their list price the same, but updating the Zestimate and status tag to reflect that.
REAL-TIME LIST PRICE TOOL
If a seller chooses the guided walkthrough, Voyager appears as a minimized sheet, just enough to stay out of the way of the content it's explaining. It walks through the mortgage balance, then preparation and repair, then closing costs, explaining the reasoning behind each estimate. It also accepts edits in dollars or a percentage. It's dismissible at any point, so a seller can take guidance on one cost and finish the rest on their own. After confirming all the costs, the next best step surfaces.
COST ESTIMATE WALKTHROUGH
The detailed Figma designs were shared with stakeholders and the broader product and engineering team at our DISCO review, where they were well received and approved. As a result, the interaction patterns and behaviors will be used across the entire Plan Tab experience.
From there, the next step was feasibility work with engineering on more complex behaviors, like auto-progressing accordions and expand and collapse of individual modules. Engineering confirmed it was feasible, the open question was timeline, whether it would land for MVP or as a fast follow.
Unfortunately, this is where my involvement ended. The day after, I was laid off, along with my teammate who'd executed the home value flow and a significant number of other designers, part of a round of cost reductions that affected 500 people company-wide. I'm confident the design was in a strong enough place for the remaining team to finish the work. I wish them much success.
“One thing I could always count on was that if you handed something to Jasmine, she'd take ownership of it and deliver it with care.”
“Whether it was a complex design problem or a tight deadline, she consistently produced thoughtful, high-quality work.”
What this work shaped
Impact
NEXT READ